OPENAI has admitted its rogue bots leaked dozens of images from ChatGPT users and meddled with US government agencies.
The worrying development comes after the AI giant revealed another accidental hacking – and follows a string of stark warnings from tech titans.
On Friday OpenAI said its agents had leaked 53 pictures from ChatGPT users.
But the tech firm did not say whether these images were AI-generated or real photos, and they did not disclose when the breach took place.
On the same day, OpenAI confirmed its rogue agents had also accessed US government websites.
Targets included the Security and Exchange Commission as well as the commerce department.
The ChatGPT maker is also probing a botched breach of the education department’s website, the New York Times reported.
It comes after a series of stark warnings from top AI CEOs that humanity could be threatened by their bots if there aren’t proper safeguards in place.
Experts have also said that rapidly advancing AI bots could break out of test labs and unleash a devastating cyberhack on the British government.
And on Wednesday Australia’s prime minister Anthony Albanese said OpenAI agents broke into a government health data portal in June.
Donald Trump downplayed warnings of AI threats as a “hoax” and ruled out US regulation.
But Albanese and other leaders reemphasised calls for greater oversight.
Earlier this year, OpenAI also admitted its bots had hacked into another tech firm Hugging Face.
The recent disclosures highlight the breakneck speed at which AI is progressing as its creators try to oversee all of the risks.
In September, a well-placed source said OpenAI had found roughly two dozen cases of agents acting in malicious ways, the Guardian reports.
But fears are mounting the true number could be much greater.
OpenAI has said a review will take months to complete due to the scale of the work.
The company said it had notified “dozens” of third parties about improper activity.
Most of the leaked images have been removed and other sites are being urged to take down the rest, OpenAI said.
Employees and researchers said OpenAI’s agents had access to the leaked images because the firm relies on anonymous user data for part of its model-training process.
ChatGPT users must deliberately opt out if they want to bar the firm from taking their data for training their large language models.
Before users’ posts are taken for training, they go through an anonymisation process.
This strips out their metadata, names, and any other information that would make it easy for the picture to be traced back to a person, according to OpenAI.
But several people close to the firm have said that this practice is not 100 per cent foolproof – and that data can still remain attached to images.